AI does use water, but there is no reliable universal figure for “water per prompt.” Estimates depend on the workload, the data center’s cooling system and location, and whether the calculation includes only on-site cooling or also electricity generation and chip manufacturing. A model-training estimate and a production measurement for a single type of prompt describe different things.
How much water does AI use?
The clearest published figures show why a single answer would mislead. A 2023 preprint estimated water used to train GPT-3 in specified U.S. data centers; a 2025 Google study measured the median text prompt served by Gemini Apps in Google’s production infrastructure. Neither figure represents every AI model or use case, and they are not a controlled comparison.
As an Amazon Associate I earn from qualifying purchases.
| Estimate | What it covers | How to interpret it |
|---|---|---|
| 700,000 liters of direct freshwater consumption | The 2023 preprint’s estimate for training GPT-3 in Microsoft’s U.S. data centers. | A modeled training estimate, not an audited measurement. The authors estimate it would be three times higher if the training took place in Microsoft’s Asian data centers. Read the 2023 preprint. |
| 0.26 milliliters of water and 0.24 watt-hours of energy | Google researchers’ 2025 measurement for a median Gemini Apps text prompt in their production-serving infrastructure, using the study’s defined functional unit. | A measurement for that service and prompt category, not a general rate for all AI prompts. Read the 2025 study or its technical report. |
The figures answer different questions: one estimates water consumed during model training, while the other measures water associated with serving a particular type of prompt. Their boundaries and methods differ, so dividing one by the other or treating them as a model-versus-model comparison would not be meaningful.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Where does AI’s water footprint come from?
Water can enter the footprint at several stages. The OECD’s 2022 review identifies data-center cooling, electricity production, and semiconductor fabrication as distinct pathways.
#1 Best Overall
Cooling data centers
Some facilities use water directly to manage heat, including evaporative or adiabatic cooling systems. Others rely on air cooling, which Google’s 2025 study says can use little to no water. Evaporative and adiabatic systems tend to consume more water during hotter, drier periods, so the same computing workload can have different direct water impacts depending on its site and operating conditions.
Generating electricity
Power generation can use water, creating an indirect footprint when data centers draw electricity from the grid. Whether that amount appears in an estimate depends on its accounting boundary. Adding an electricity-related estimate to a direct-cooling estimate is only sound if the two do not overlap.
Manufacturing chips
Semiconductor fabrication is another part of the wider footprint. A figure limited to operating a data center may omit manufacturing, just as an estimate limited to chip production would omit the water involved in running servers. Check what a source includes before comparing its total with another estimate.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why can the same workload have different water estimates?
A prompt does not have a fixed water cost independent of where and how it is served. Key variables include the facility’s cooling design, weather, location, computing infrastructure, and the electricity supply. Estimates also vary with the task being measured: training a model, processing work in batches, and answering an interactive prompt are different workloads.
- Workload: Is the estimate for training, batch inference, or interactive use—and what model or task does it cover?
- Boundary: Does it count direct cooling, electricity-related water, chip manufacturing, or some combination?
- Water measure: Is the figure for withdrawal, consumption, or replenishment?
- Place and time: Which facility, climate, season, and reporting period does it describe?
- Method: Is it an instrumented production measurement, a company-wide report, a model-based estimate, or a scenario?
Google’s 2025 study warns that site and seasonal differences can distort comparisons. For comparing data-center fleets, it recommends annual or trailing-twelve-month averages rather than a snapshot from one location or season.
What do company water totals tell us—and what don’t they?
Company-wide or fleet-level numbers provide context, but they do not isolate AI’s share. Google reported that its data centers consumed 6.1 billion gallons of water in 2023, 17% more than in 2022. The report said the increase mirrored growth in electricity use; it does not establish that AI alone caused the rise. Google’s 2024 Environmental Report also said the company replenished 1 billion gallons, equal to 18% of its 2023 freshwater consumption. Replenishment is reported separately and does not mean consumption fell at a particular data center.
Microsoft’s current sustainability report says it replenished 14.2 million cubic meters in FY25. That is a company-level replenishment figure, not an AI-attributed water-use total. See Microsoft’s Environmental Sustainability Report.
Recommended Free Tools
Local context matters as well as corporate totals. An Associated Press report published September 9, 2023, described Microsoft data-center operations in Iowa and reported that Microsoft’s company-wide water consumption rose 34% from 2021 to 2022. The story quoted researcher Shaolei Ren attributing most of that growth to AI; that was his assessment, not a company-published measurement of AI’s share. Ren told AP, “Most people are not aware of the resource usage underlying ChatGPT.” Read AP’s report.
Best Value
What do withdrawal, consumption, and replenishment mean?
These terms describe different things, so they should not be treated as interchangeable. Withdrawal is water taken from a source. Consumption is water not returned to the same source in a way that is available for use, such as water lost through evaporation. Replenishment describes efforts to restore or benefit water resources; it is not the same measurement as water consumed at an individual site.
The OECD’s 2022 review said water accounting for data centers was less developed than accounting for energy use and greenhouse-gas emissions. Citing earlier sources, it estimated that 33–50% of data-center operators compiled and reported water-use metrics. That is a historical estimate from the report, not a current census of operators. The OECD also noted that even when a data center’s share of national water use is small, it can compete with other users in a water-scarce location.
What would make AI’s water footprint easier to assess?
Providers and operators can make estimates more useful by disclosing the workload, accounting boundary, water metric, location, reporting period, and measurement method. Efficiency choices, site selection, cooling design, and consistent monitoring also affect facility water use. Without comparable disclosures, a striking per-prompt number or corporate total should be read only within the scope its source actually measured.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
